CN107077625A - Hierarchical Deep Convolutional Neural Networks - Google Patents

Abstract

The deep convolutional neural networks (HD CNN) of layering branch improve existing convolutional neural networks (CNN) technology.In HD CNN, the class that can be easily distinguished is classified in high-rise thick classification CNN, and being sorted in lower floor fine classification CNN for being most difficult to is completed.In HD CNN training, multinomial logistics loss and novel time degree of rarefication punishment can be used.The using of multinomial logistics loss and the punishment of the time degree of rarefication classification subset that to cause each branch component processing different.

Description

Hierarchical Deep Convolutional Neural Networks

priority statement

This application claims priority to U.S. Patent Application No. 14/582,059, filed December 23, 2014, entitled "Hierarchical Deep Convolutional Neural Network For Image Classification," which claims The priority of US Patent Application No. 62/068,883 for "Hierarchical Deep Convolutional Neural Network For Image Classification," each of which is incorporated herein by reference.

technical field

The subject matter disclosed herein generally involves classifying data using layered deep convolutional neural networks. In particular, the present disclosure relates to systems and methods for generating and using hierarchical deep convolutional neural networks for image classification.

Background technique

A deep convolutional neural network (CNN) is trained as an N-way classifier to distinguish between N classes of data. CNN classifiers are used to classify images, detect objects, estimate poses, recognize faces, and perform other classification tasks. Usually, the structure of CNN (eg, number of layers, type of layers, connectivity between layers, etc.) is selected by the designer, and then the parameters of each layer are determined through training.

Multiple classifiers can be combined by averaging. In model averaging, multiple individual models are used. Each model is capable of classifying the entire set of categories, and each model is trained independently. The main sources of variance in its predictions include: different initializations, different subsets of the training corpus, etc. The output of the combined model is the average output of the individual models.

Description of drawings

Some embodiments are shown in the figures by way of example and not limitation.

FIG. 1 is a network diagram illustrating a network environment suitable for creating and using a hierarchical deep CNN for image classification, according to some example embodiments.

FIG. 2 is a block diagram illustrating components of a hierarchical deep CNN server suitable for image classification, according to some example embodiments.

FIG. 3 is a block diagram illustrating components of an apparatus suitable for image classification using hierarchical deep CNN techniques, according to some example embodiments.

FIG. 4 is a group diagram illustrating classification images according to some example embodiments.

FIG. 5 is a block diagram illustrating a relationship between components of a server configured to recognize fine-grained categories of images, according to some example embodiments.

FIG. 6 is a flowchart illustrating an operation of a server in performing a process of recognizing a coarse category, according to some example embodiments.

FIG. 7 is a flowchart illustrating the operation of a server in performing a process of generating a hierarchical deep CNN for classifying images, according to some example embodiments.

Figure 8 is a block diagram illustrating an example of a software architecture that may be installed on a machine according to some example embodiments.

9 shows a diagrammatic representation of a machine in the form of a computer system in which a set of instructions can be executed to cause the machine to perform any one or more of the methods discussed herein, according to an example embodiment. method.

detailed description

Example methods and systems relate to hierarchical deep CNNs for image classification. The examples merely represent possible variations. Components and functions are optional and may be combined or subdivided, and operations may be changed in order or combined or subdivided, unless explicitly stated otherwise. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of example embodiments. It will be apparent, however, to one skilled in the art that the present subject matter may be practiced without these specific details.

Hierarchical Deep CNN (HD-CNN) follows a coarse-to-fine classification strategy and modular design principles. For any given class label, it is possible to define a collection of simple classes and a collection of confused classes. Therefore, the initial coarse classifier CNN is able to separate easily separable classes from each other. Subsequently, the challenging classes are routed to a downstream fine CNN, which only focuses on the confused classes. In some example embodiments, HD-CNN improves classification performance over standard deep CNN models. Like CNN, the structure of HD-CNN (e.g., the structure of each component CNN, the number of fine classes, etc.) can be determined by the designer, while the parameters of each layer of each CNN can be determined by training.

Compared with training HD-CNN from scratch, pre-training HD-CNN can gain benefits. For example, compared to standard deep CNN models, HD-CNN has additional free parameters from shared branch shallow layers as well as C′ independent branch deep layers. This will greatly increase the number of free parameters within HD-CNN compared to standard CNN. Therefore, overfitting is more likely to occur in HD-CNN if the same amount of training data is used. Pre-training can help overcome the difficulty of insufficient training data.

Another potential benefit of pre-training is that a good choice of coarse categories will benefit training branch components to focus on a consistent subset of confusingly fine categories. For example, Branch 1 is good at distinguishing apples from oranges, while Branch 2 is better at distinguishing buses from trains. Thus, a coarse set of categories is identified and the coarse category component is pre-trained to classify the set of coarse categories.

Some training datasets include information about coarse categories and the relationship between fine and coarse categories. However, this is not the case for many training datasets. These training datasets only provide fine categories for each item in the dataset and do not identify coarse categories. Therefore, a process of classifying fine categories into coarse categories is described below with reference to FIG. 6 .

FIG. 1 is a network diagram illustrating a network environment 100 suitable for creating and using a hierarchical deep CNN for image classification, according to some example embodiments. Network environment 100 includes e-commerce servers 120 and 140 , HD-CNN server 130 , and devices 150A, 150B, and 150C, which are all communicatively coupled to one another via network 170 . Devices 150A, 150B, and 150C may be referred to collectively as "device 150," or collectively as "device 150." E-commerce servers 120 and 140 and HD-CNN server 130 may be part of web-based system 110 . Alternatively, the device 150 may be directly connected to the HD-CNN server 130 or connected to the HD-CNN server 130 through a local network different from the network 170 used to connect to the e-commerce server 120 or 140 . As described below with reference to FIGS. 8-9 , the e-commerce servers 120 and 140 , the HD-CNN server 130 , and the device 150 may be wholly or partially implemented in a computer system.

E-commerce servers 120 and 140 provide e-commerce applications to other machines (eg, device 150 ) via network 170 . The e-commerce servers 120 and 140 can also be directly connected to the HD-CNN server 130 or integrated with the HD-CNN server 130 . In some example embodiments, one e-commerce server 120 and HD-CNN server 130 are part of web-based system 110 , while other e-commerce servers (eg, e-commerce server 140 ) are separate from web-based system 110 . An e-commerce application may provide a way for users to buy and sell items directly to each other, to buy items from and sell items to an e-commerce application provider, or both.

The HD-CNN server 130 creates an HD-CNN for classifying images, uses the HD-CNN to classify images, or performs both. For example, HD-CNN server 130 may create an HD-CNN for classifying images based on the training set, or may load a pre-existing HD-CNN onto HD-CNN server 130 . The HD-CNN server 130 may also respond to requests for image classification by providing fine-grained categories for the images. HD-CNN server 130 may provide data to other machines (eg, e-commerce servers 120 and 140 or device 150 ) via network 170 or another network. HD-CNN server 130 may receive data from other machines (eg, e-commerce servers 120 and 140 or device 150 ) via network 170 or another network. In some exemplary embodiments, the functionality of the HD-CNN server 130 described herein is performed on a user device such as a personal computer, tablet computer, or smartphone.

Also shown in FIG. 1 is user 160 . User 160 may be a human user (e.g., a human), a machine user (e.g., a computer configured by a software program to interact with device 150 and HD-CNN server 130), or any suitable combination thereof (e.g., a machine-assisted human or human supervised machine). User 160 is not part of network environment 100 , but is associated with device 150 and may be a user of device 150 . For example, device 150 may be a sensor belonging to user 160 , a desktop computer, a vehicle computer, a tablet computer, a navigation device, a portable media device, or a smart phone.

In some exemplary embodiments, HD-CNN server 130 receives data related to items of interest to the user. For example, a camera attached to device 150A may take an image of an item that user 160 wishes to sell and send the image to HD-CNN server 130 over network 170 . HD-CNN server 130 classifies the item based on the image. Categories may be sent to e-commerce server 120 or 140, to device 150A, or any combination thereof. The e-commerce server 120 or 140 may use this category to assist in generating a list of items for sale. Similarly, the image may be of an item of interest to user 160 , and the categories may be used by e-commerce server 120 or 140 to help select a list of items to display to user 160 .

Any of the machines, databases, or devices shown in Figure 1 can be implemented with a general-purpose computer that is modified (e.g., configured or programmed) by software into a special-purpose computer to perform the tasks described herein for the machines, databases, or The capabilities described by the device. For example, a computer system capable of implementing any one or more of the methods described herein is discussed below with reference to FIGS. 8-9 . As used herein, a "database" is a data storage resource and can store data structured as text files, tables, spreadsheets, relational databases (e.g., object-relational databases), triple stores, hierarchical data storage or any suitable combination thereof. Furthermore, any two or more of the machines, databases, or devices shown in Figure 1 may be combined into a single machine, and the functionality described herein for any single machine, database, or device may be subdivided into multiple machines, databases, or or in the device.

Network 170 may be any network that enables communication between machines, databases, and devices (eg, HD-CNN server 130 and device 150). Thus, network 170 may be a wired network, a wireless network (eg, a mobile or cellular network), or any suitable combination thereof. Network 170 may include one or more components that constitute a private network, a public network (eg, the Internet), or any suitable combination thereof.

FIG. 2 is a block diagram illustrating components of the HD-CNN server 130 according to some example embodiments. HD-CNN server 130 is shown to include a communication module 210, a coarse class recognition module 220, a pre-training module 230, a fine-tuning module 240, a classification module 250, and a storage module 260, all of which are configured to communicate with each other (e.g., via a bus , shared storage, or switches). Any one or more of the modules described herein can be implemented using hardware, such as a processor of a machine. Furthermore, any two or more of these modules may be combined into a single module, and functions described herein for a single module may be subdivided into multiple modules. Furthermore, modules described herein as being implemented in a single machine, database or device may be distributed among multiple machines, databases or devices according to various example embodiments.

The communication module 210 is configured to send and receive data. For example, the communication module 210 may receive image data through the network 170 and transmit the received data to the classification module 250 . As another example, the classification module 250 may identify the category of the item, and the communication module 210 may send the category of the item to the e-commerce server 120 through the network 170 .

The coarse category identification module 220 is configured to identify coarse categories for a given data set. The coarse category identification module 220 determines related fine categories and groups them into coarse categories. For example, a provided dataset may have C fine categories, and the HD-CNN designer may determine the desired number of coarse categories C'. The coarse category identification module 220 identifies a mapping of C fine categories to C' coarse categories. Fine categories may be grouped into coarse categories using process 600 of FIG. 6 described below.

The pre-training module 230 and the fine-tuning module 240 are configured to determine parameters of the HD-CNN. The pre-training module 230 pre-trains the coarse category CNN and the fine category CNN to reduce the overlap between the fine category CNNs. After the pre-training is complete, the fine-tuning module 240 provides additional adjustments to the HD-CNN. Pre-training and fine-tuning may be performed using process 700 of FIG. 7 described below.

Classification module 250 is configured to receive and process image data. The image data may be two-dimensional images, frames from a continuous video stream, three-dimensional images, depth images, infrared images, binocular images, or any suitable combination thereof. For example, an image may be received from a camera. To illustrate, a camera may take a picture and send it to the classification module 250 . The classification module 250 determines a fine class of an image by using an HD-CNN (eg, by using a coarse class CNN to determine a coarse class or a coarse class weight, and using one or more fine class CNNs to determine a fine class). The HD-CNN can be generated using the pre-training module 230, the fine-tuning module 240, or both. Alternatively, HD-CNN may be provided from an external source.

The storage module 260 is configured to store and retrieve data generated and used by the coarse category identification module 220 , the pre-training module 230 , the fine-tuning module 240 and the classification module 250 . For example, the HD-CNN generated by the pre-training module 230 may be stored by the storage module 260 for retrieval by the fine-tuning module 240 . Information on image classification generated by the classification module 250 may also be stored by the storage module 260 . E-commerce server 120 or 140 may request (eg, by providing an image identifier) a category of images that may be retrieved from memory by storage module 260 and sent over network 170 using communication module 210 .

FIG. 3 is a block diagram illustrating components of device 150 according to some example embodiments. Device 150 is shown to include input module 310 , camera module 320 and communication module 330 all configured to communicate with one another (eg, via a bus, shared memory, or switch). Any one or more of the modules described herein can be implemented using hardware, such as a processor of a machine. Furthermore, any two or more of these modules may be combined into a single module, and functions described herein for a single module may be subdivided into multiple modules. Furthermore, modules described herein as being implemented in a single machine, database or device may be distributed among multiple machines, databases or devices according to various example embodiments.

The input module 310 is configured to receive input from a user via a user interface. For example, a user may enter their username and password into the input module, configure the camera, select an image to use as the basis for a list or item search, or any suitable combination thereof.

The camera module 320 is configured to capture image data. For example, an image may be received from a camera, a depth image may be received from an infrared camera, a pair of images may be received from a binocular camera, and so on.

The communication module 330 is configured to transmit data received by the input module 310 or the camera module 320 to the HD-CNN server 130 , the e-commerce server 120 or the e-commerce server 140 . For example, input module 310 may receive a selection of an image captured with camera module 320 and an indication that the image depicts an item that a user (eg, user 160 ) wishes to sell. The communication module 330 may transmit the images and instructions to the e-commerce server 120 . The e-commerce server 120 may send the image to the HD-CNN server 130 to request classification of the image, generate a list template based on the category, and cause the list template to be presented to the user via the communication module 330 and the input module 310 .

FIG. 4 is a group diagram illustrating classification images according to some example embodiments. In FIG. 4, twenty seven images have been correctly classified as depicting apples (group 410), oranges (group 420), or buses (group 430). Groups 410-430 are referred to herein as apples, oranges and buses. By inspection, it is relatively easy to distinguish members of apples from members of buses, and more difficult to distinguish members of apples from members of oranges. Images from apples and oranges may have similar shapes, textures, and colors, so distinguishing them correctly is harder. In contrast, images from buses often have a different visual appearance than apples, so easier classification can be expected. In fact, the two categories of apples and oranges can be defined as belonging to the same coarse category, while buses belong to different coarse categories. For example, in the CIFAR 100 dataset (discussed in "Learning Multiple Layers of Features from Tiny Images", Krizhevsky (2009)), apples and oranges are subcategories in "Fruits and Vegetables", while buses are subcategories in " subcategory in Vehicle 1". The CIFAR 100 dataset consists of 100 classes of natural images. There are 50,000 training images and 10,000 testing images in the CIFAR 100 dataset.

FIG. 5 is a block diagram illustrating the relationship between components of the classification module 250 according to some example embodiments. A single standard deep CNN can be used as a building block for the fine prediction component of HD-CNN. As shown in Figure 5, the coarse class CNN 520 predicts probabilities on the coarse class. Multiple branch CNNs 540-550 are added independently. In some exemplary embodiments, branched CNNs 540 - 550 share branched shallow layer 530 . Coarse category CNN 520 and multiple branch CNNs 540-550 each receive an input image and operate on the input image in parallel. Although each branch CNN 540-550 takes an input image and gives a probability distribution over the corpus of fine categories, the results of each branch CNN 540-550 are only valid for a subset of categories. Multiple full predictions from branch CNNs 540-550 are linearly combined by a probability averaging layer 560 to form a final fine class prediction weighted by the corresponding coarse classification probabilities.

The following symbols are used in the discussion below. The data set includes: N _t training samples { _xi , y _i } ^t (where i is in the range of 1 to N _t ), and N _s test samples { _xi , y _i } ^t (where i is in the range of 1 to N t N _s range). x _i and y _i denote image data and image labels, respectively. Image labels correspond to fine-grained categories of images. There are C predefined fine categories in the dataset {S _k }, where k ranges from 1 to C. There are C' coarse categories in this dataset.

Like standard deep CNN models, HD-CNN achieves end-to-end classification. Although the standard deep CNN model consists of only a single CNN, HD-CNN mainly includes three parts, namely, a single coarse category component B (corresponding to the coarse category CNN 520), multiple branch fine category components {F ^j } (where, j in the range of 1 to C' (corresponding to branch CNN 540-550)), and a single probability averaging layer (corresponding to probability averaging layer 560). A single coarse category CNN 520 receives as input raw image pixel data and outputs a probability distribution over the coarse category. The coarse class probabilities are used by the probability averaging layer 560 to assign weights to the full predictions made by the branch CNNs 540-550.

Figure 5 also shows a collection of branch CNNs 540-550, each branch CNN making over-corpus predictions for fine categories. In some exemplary embodiments, branched CNNs 540-550 share parameters in shallow layer 530, but have independent deep layers. Shallow layers are the layers in the CNN that are closest to the original input, while deep layers are the layers that are closer to the final output. Sharing parameters in shallow layers can lead to the following benefits. First, in shallow layers, each CNN can extract raw low-level features (e.g., blobs, corners), which are useful for classifying all fine categories. Thus, shallow layers can be shared between branching components even though each branching component focuses on a different set of fine-grained categories. Second, sharing parameters in shallow layers greatly reduces the total number of parameters in HD-CNN, which can contribute to the training success of HD-CNN models. If the fine category components of each branch are trained completely independently of each other, the number of free parameters in HD-CNN will scale linearly with the number of coarse categories. An excessively large number of parameters in the model will increase the possibility of overfitting. Third, the computational cost and memory consumption of HD-CNN are also reduced due to the sharing of shallow layers, which is of practical importance for deploying HD-CNN in real applications.

Probabilistic averaging layer 560 receives all branch CNN 540-550 predictions as well as coarse category CNN 520 predictions and produces a weighted average as the final prediction p( _xi ) for image i, as shown in the following equation.

In this equation, B _ij is the probability of coarse class j for image i predicted by the coarse class CNN 520 . For image i, the fine class prediction made by the jth branch component F ^j is p _j ( _xi ).

Both the coarse category CNN 520 and the branch CNNs 540-550 can be implemented as any end-to-end deep CNN model that takes a raw image as input and returns a probability prediction over the category as output.

Using a temporal sparsity penalty on the multinomial logistic loss function used to train the fine-category component will encourage each branch to focus on a subset of the fine-category components. A modified loss function incorporating this temporal sparsity penalty is shown by the following equation:

In this equation, n is the size of the training mini-batch, _yi is the ground truth label for image i, and λ is the regularization constant. In some example embodiments, a value of 5 is used for λ. B _ij is the probability of coarse class j for image i predicted by the coarse class CNN 520 . The target temporal sparsity of branch j is denoted t _j .

Combined with branch initialization, the temporal sparsity term can ensure that each branch component focuses on classifying a different subset of fine categories and prevents a few branches from receiving a large portion of the coarse category probability body.

FIG. 6 is a flowchart illustrating the operation of the HD-CNN server 130 in performing a process 600 of identifying coarse categories, according to some example embodiments. Process 600 includes operations 610 , 620 , 630 , 640 , and 650 . By way of example only and not limitation, operations 610-650 are described as being performed by modules 210-260.

In operation 610, the coarse category identification module 220 divides the training sample set into a training set and an evaluation set. For example, divide the data set { _xi , y _i } ^t consisting of Nt training samples into two parts train_train and train_val, where i is in the range of 1 to Nt. This can be done by choosing the desired distribution of samples between train_train and train_val, for example a 70% vs. 30% distribution. Once the distribution is chosen, for each set, samples can be randomly selected in appropriate proportions. In operation 620, the deep CNN model is trained based on train_train by the pre-training module 230 using standard training techniques. For example, the back-propagation training algorithm is an option for training deep CNN models.

In operation 630, the coarse class recognition module 220 draws a confusion matrix based on train_val. The size of the confusion matrix is C×C. The columns of the matrix correspond to the predicted fine classes, while the rows of the matrix correspond to the actual fine classes in train_val. For example, if every prediction is correct, only the cells in the main diagonal of the matrix will be nonzero. Conversely, if every prediction is incorrect, then the cells in the main diagonal of the matrix will all be zero.

The coarse class identification module 220 generates the distance matrix D by subtracting each element of the confusion matrix from 1 and zeroing the diagonal elements of D. The distance matrix is made symmetric by averaging D and ^DT (the transpose of D). After performing these operations, each element D _ij measures the ease with which category i is distinguished from category j.

In operation 640, a fine-category low-dimensional feature representation {f _i } is obtained, where i is in the range of 1 to C. For example, Laplacian eigenmaps can be used for this purpose. Low-dimensional feature representations preserve local neighborhood information on low-dimensional manifolds and are used to cluster fine categories into coarse ones. In an example embodiment, the adjacency graph is constructed using k nearest neighbors. For example, a value of 3 could be used for k. The weights of the adjacency graph are set by using a hot kernel (eg, with a width parameter t=0.95). In some exemplary embodiments, {f _i } has a dimension of three.

The coarse category identification module 220 clusters (in operation 650) the C fine categories into C' coarse categories. Clustering may be performed using affinity propagation, k-means clustering, or other clustering algorithms. Affine propagation can automatically introduce the number of coarse categories and may lead to clusters that are more balanced in size compared to other clustering methods. Balanced clustering helps ensure that each branch component handles a similar number of fine-grained categories and thus has a similar workload. The damping factor λ in the affine propagation may affect the number of resulting clusters. In some exemplary embodiments, λ is set to 0.98. The result of the clustering is the mapping P(y)=y' from the fine class y to the coarse class y'.

For example, the 100 categories of the CIFAR100 dataset can be divided into coarse categories by training a deep CNN model based on the dataset's 50,000 training images and 10,000 testing images. The number of coarse categories may be provided as input (eg, four coarse categories may be selected), and process 600 is used to divide the fine categories into coarse categories. In an example embodiment, the 100 categories of the CIFAR100 dataset are divided into four coarse categories, as shown in the table below.

FIG. 7 is a flowchart illustrating the operation of the HD-CNN server 130 in performing a process 700 of generating an HD-CNN for classifying images, according to some example embodiments. Process 700 includes operations 710 , 720 , 730 , 740 , 750 , and 760 . By way of example only and not limitation, operations 710-760 are described as being performed by modules 210-260.

In operation 710, the pre-training module 230 trains a coarse category CNN on the set of coarse categories. For example, a set of coarse categories may have been identified using process 600 . Using the mapping P(y) = y', the fine categories of the training dataset are replaced by the coarse categories. In an example embodiment, the dataset { _xi , _y'i } is used to train a standard deep CNN model, where i ranges from 1 to Nt. The trained model becomes the coarse category component of HD-CNN (eg, coarse category CNN520).

In an example embodiment, a network consisting of three convolutional layers, one fully connected layer and one SOFTMAX layer is used. Each convolutional layer has 64 filters. Rectified linear units (ReLU) are used as activation units. Pooling and response normalization layers are also used between convolutional layers. The complete example schema is defined in the Example 1 table below. Another example schema is defined in the Example 2 table below.

In the table above, the filter uses the indicated number of inputs (eg, pixel values). For example, a 5x5 filter looks at 25 pixels in a 5x5 grid to determine a single value. The 5x5 filter considers every 5x5 grid in the input image. Thus, a layer with 64 5x5 filters generates 64 outputs for each input pixel, each of these values based on a 5x5 pixel grid centered on that input pixel. MAX pooling has multiple inputs for a collection of pixels and provides a single output, the maximum of those inputs. For example, a 3x3MAX pooling layer will output a value for each 3x3 pixel block, which is the maximum value of those 9 pixels. AVG pooling has multiple inputs for sets of pixels and provides a single output, the average (eg mean) of those inputs. A normalization layer normalizes the values output from the previous layer. The cccp layer provides nonlinear components to the CNN. The SOFTMAX function is a normalized exponential function that provides a non-linear variant of multinomial stream regression. In some example embodiments, the SOFTMAX function takes a K-dimensional value vector and outputs a K-dimensional value vector such that the sum of the elements of the output vector is 1 and ranges from 0 to 1. For example, the following equation can be used to generate an output vector y from an input vector z:

where j=1,...,K.

In operation 720, the pre-training module 230 also trains the prototype fine category component. For example, the dataset { _xi , _yi } (where i ranges from 1 to Nt) is used to train a standard deep CNN model, which becomes the prototypical fine-category component. In an example embodiment, the CIFAR100 dataset is used to train a CNN as a prototype fine category component.

In operation 730, a loop begins to process each of the C' fine-category components. Accordingly, operations 740 and 750 are performed for each fine category component. For example, when four coarse categories are identified, the loop iterates for each of the four fine category components.

In operation 740, the pre-training module 230 makes a copy of the prototype fine-category component of the fine-category component. Therefore, all fine-category components are initialized to the same state. The fine category component is further trained on the portion of the dataset corresponding to the coarse category of the fine category component. For example, a subset of the dataset {x _i , y _i } can be used, where P(yi) is the coarse category. Once all fine and coarse category components have been trained, the HD-CNN is constructed.

The shallow layers of a CNN targeting fine category components can be kept fixed, while the deep layers are allowed to change during training. For example, using the structure of example 1 above, for each fine class component, the shallow conv1, pool1 and norm1 can remain unchanged, while the deep conv2, pool2, norm2, conv3, pool3, ip1 and prob are in each fine class component Modified during training. In some exemplary embodiments, the structure of the shallow layer remains fixed, but the values used within the shallow layer are allowed to change. Regarding the structure of Example 2 above, for each fine category component, the shallow conv1, cccp1, cccp2, pool1 and conv2 can remain unchanged, while the deep cccp3, cccp4, pool2, conv3, cccp5, cccp6, pool3 and prob are in each The fine category component is modified during training.

In operation 760, the fine-tuning module 240 fine-tunes the constructed HD-CNN. Fine-tuning can be performed using a multinomial logistic loss function with a temporal sparsity penalty. A target temporal sparsity {tj} (where j ranges from 1 to C') can be defined using a map P. For example, the following equation can be used, where S _k is the set of images from fine category k.

The batch size for fine-tuning can be chosen based on computation time and desired amount of learning per iteration. For example, a batch size of 250 can be used. After each batch, the training error can be measured. If the rate of improvement in training error is below a threshold, the learning rate may be reduced (eg, by 10%, by half, or by another amount). The threshold can be modified when the learning rate is reduced. The fine-tuning process stops after the minimum learning rate is reached (eg, when the learning rate drops below 50% of the original value), after a predetermined number of batches have been used for fine-tuning, or any suitable combination thereof.

According to various example embodiments, one or more of the methods described herein may facilitate generation of an HD-CNN for image classification. Furthermore, one or more of the methods described in this paper can facilitate the classification of images with higher success rates relative to standard deep CNNs. Furthermore, one or more of the methods described herein may facilitate training HD-CNNs for users more quickly and using less computing power than previous methods. Similarly, one or more of the methods described herein may facilitate training an HD-CNN at the same resolution quality with fewer training samples than is the case with training a CNN to the same resolution quality.

When these effects are considered collectively, one or more of the methods described herein can eliminate the need for some of the workload or resources that would otherwise be generated or used by an HD-CNN for image classification. Involved. By one or more of the methods described herein, user effort in ordering an item of interest may also be reduced. For example, accurately identifying the category of items a user is interested in based on an image can reduce the time or effort a user spends creating a list of items or finding items to purchase. Computing resources used by one or more machines, databases, or devices (eg, in network environment 100) may similarly be reduced. Examples of such computing resources include processor cycles, network traffic, memory usage, data storage capacity, power consumption, and cooling capabilities.

Software Architecture

FIG. 8 is a block diagram 800 illustrating the architecture of software 802 that may be installed on any one or more of the devices described above. Figure 8 is only a non-limiting example of a software architecture, and it should be appreciated that many other architectures can be implemented to facilitate the functionality described herein. Software 802 may be implemented by hardware such as machine 900 of FIG. 9 , which includes processor 910 , memory 930 , and I/O components 950 . In this example architecture, software 802 can be conceptualized as a stack of layers, where each layer can provide specific functionality. For example, software 802 includes layers such as operating system 804 , libraries 806 , framework 808 , and applications 810 . Operationally, according to some embodiments, an application 810 invokes an application programming interface (API) call 812 through a software stack and receives a message 814 in response to the API call 812 .

In various implementations, the operating system 804 manages hardware resources and provides common services. Operating system 804 includes, for example, kernel 820 , services 822 and drivers 824 . In some implementations, core 820 acts as an abstraction layer between hardware and other software layers. For example, kernel 820 provides functions for memory management, processor management (eg, scheduling), component management, networking and security settings, among others. Services 822 may provide other common services to other software layers. The driver 824 may be responsible for controlling or interfacing with the underlying hardware. For example, drivers 824 may include display drivers, camera drivers, drivers, flash drives, serial communication drivers (such as Universal Serial Bus (USB) drivers), drivers, audio drivers, power management drivers, and more.

In some implementations, libraries 806 provide low-level common infrastructure that can be used by applications 810 . Libraries 806 may include system libraries 830 (eg, a C standard library) that may provide functionality such as memory allocation functions, string manipulation functions, math functions, and the like. Additionally, libraries 806 may include API libraries 832, such as media libraries (e.g., libraries that support rendering and manipulation of various media formats, such as Moving Picture Experts Group 4 (MPEG4), Advanced Video Coding (H.264 or AVC), Moving Picture Experts Group Layer 3 (MP3), Advanced Audio Coding (AAC), Adaptive Multi-Rate (AMR) audio codec, Joint Photographic Experts Group (JPEG or JPG), or Portable Network Graphics (PNG)) , graphics libraries (such as the OpenGL framework for two-dimensional (2D) and three-dimensional (3D) rendering in graphical content on a display), databases (such as SQLite that provides various relational database functions), web libraries (such as , WebKit that provides web browsing functionality), etc. Library 806 may also include a variety of other libraries 834 to provide application 810 with many other APIs.

According to some implementations, framework 808 provides a high-level common infrastructure that can be used by applications 810 . For example, framework 808 provides various graphical user interface (GUI) functions, advanced resource management, advanced location services, and the like. Framework 808 may provide a wide range of other APIs that may be used by applications 810, some of which may be specific to a particular operating system or platform.

In an example embodiment, applications 810 include home application 850, contacts application 852, browser application 854, book reader application 856, location application 858, media application 860, messaging application 862, gaming application 864, and third-party applications such as Various other applications such as application 866. According to some embodiments, the application 810 is a program that performs functions defined in the program. Various programming languages may be employed to create one or more of the variously structured applications 810, such as object-oriented programming languages (e.g., Objective-C, Java, or C++) or procedural programming languages (e.g., C or assembly language). In a specific example, a third-party application 866 (e.g., an application developed by an entity other than the vendor of the particular platform using the ANDROID™ or IOS ^™ software development kit (SDK)) may be running on a mobile operating system such as iOS ^™ , Android ^TM , Mobile software running on the Phone or other mobile operating system). In this example, third party application 866 can invoke API calls 812 provided by mobile operating system 804 to facilitate the functionality described herein.

Example machine architecture and machine-readable media

9 is a diagram illustrating components of a machine 900 capable of reading instructions from a machine-readable medium (eg, a machine-readable storage medium) and performing any one or more of the methods discussed herein, according to some example embodiments block diagram. In particular, FIG. 9 shows a schematic representation of a computer system in the form of a machine 900 in which instructions 916 (e.g., software, programs, applications, applets, apps, or other executable code) can be executed. to cause machine 900 to perform any one or more of the methods discussed herein. In alternative embodiments, machine 900 operates as a standalone device or may be coupled (eg, networked) to other machines. In a networked deployment, the machine 900 can operate in the capacity of a server machine or a client machine in a server-client network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. Machine 900 may include, but is not limited to, server computers, client computers, personal computers (PCs), tablet computers, laptop computers, netbooks, set-top boxes (STBs), personal digital assistants (PDAs), entertainment media systems, cellular phones, smart Phones, mobile devices, wearable devices (such as smart watches), smart home devices (such as smart appliances), other smart devices, network devices, network routers, network switches, network bridges, or machines capable of sequentially or otherwise executing specified 900 any machine that instructs 916 an action to take. Further, while a single machine 900 is shown, the term "machine" will also be taken to include a collection of machines 900 that individually or jointly execute instructions 916 to perform any one or more of the methodologies discussed herein.

Machine 900 may include a processor 910 , memory 930 , and I/O components 950 that may be configured to communicate with one another via a bus 902 . In an example embodiment, the processor 910 (eg, central processing unit (CPU), reduced instruction set computing (RISC) processor, complex instruction set computing (CISC) processor, graphics processing unit (GPU), digital signal processor (DSP, application specific integrated circuit (ASIC), radio frequency integrated circuit (RFIC), other processors, or any suitable combination thereof) may include, for example, processor 912 and processor 914 that may execute instructions 916 . The term "processor" is intended to include multi-core processors, which may include two or more independent processors (also referred to as "cores") that can execute instructions concurrently. Although FIG. 9 shows multiple processors, machine 900 may include a single processor with a single core, a single processor with multiple cores (e.g., multi-core processing), multiple processors with a single core, multiple processors with multiple cores or any combination thereof.

The memory 930 may include a main memory 932 , a static memory 934 and a storage unit 936 accessible by the processor 910 via the bus 902 . The storage unit 936 may include a machine-readable medium 938 having stored thereon instructions 916 implementing any one or more of the methods or functions described herein. Instructions 916 may also reside, completely or at least partially, within at least one of main memory 932, static memory 934, processor 910 (e.g., within a cache memory of the processor), during execution of the instructions by machine 900, or any suitable combination thereof. Accordingly, main memory 932 , static storage 934 and processor 910 are considered machine-readable media 938 in various implementations.

As used herein, the term "memory" refers to a machine-readable medium 938 capable of storing data temporarily or permanently, and may be considered to include, but is not limited to, random access memory (RAM), read only memory (ROM), cache memory, flash memory, and cache memory. Although machine-readable medium 938 is shown as a single medium in example embodiments, the term "machine-readable medium" should be taken to include a single medium or multiple media capable of storing instructions 916 (e.g., a centralized or distributed database or associated cache and server). The term "machine-readable medium" will also be taken to include any medium or combination of media capable of storing instructions (e.g., instructions 916) to be executed by a machine (e.g., machine 900) such that the instructions are executed by one or more of the machine 900 A processor (eg, processor 910), when executed, causes machine 900 to perform any one or more of the methods described herein. Accordingly, a "machine-readable medium" refers to a single storage device or device, as well as to a "cloud-based" storage system or storage network that includes multiple storage devices or devices. Accordingly, the term "machine-readable medium" should be understood to include, but is not limited to, devices with solid-state memory (e.g., flash memory), optical media, magnetic EPROM)) or any suitable combination thereof etc. in the form of one or more data repositories. The term "machine-readable medium" specifically excludes non-statutory signals per se.

I/O components 950 include various components for receiving input, providing output, generating output, sending information, exchanging information, capturing measurements, and the like. In general, it should be understood that I/O components 950 may include many other components not shown in FIG. 9 . I/O components 950 may be grouped according to function only to simplify the following discussion, and the grouping is not limiting in any way. In various example embodiments, I/O components 950 include output components 952 and input components 954 . Output components 952 include visual components (e.g., a display such as a plasma display panel (PDP), light emitting diode (LED) display, liquid crystal display (LCD), projector, or cathode ray tube (CRT)), acoustic components (e.g., a speaker ), haptic components (eg, vibration motors), other signal generators, etc. Input components 954 include alphanumeric input components (e.g., a keyboard, a touch screen configured to receive alphanumeric input, an optical keyboard, or other alphanumeric input components), point-based input components (e.g., a mouse, touchpad, sticks, motion sensors, or other pointing instruments), tactile input components (e.g., physical buttons, touch screens that provide the position and force of touch or touch gestures, or other tactile input components), audio input components (e.g., microphones), etc.

In other example embodiments, the I/O component 950 includes components such as a biometric component 956, a motion component 958, an environmental component 960, or a location component 962, among others. For example, the biometric component 956 includes functions for detecting manifestations (e.g., hand expression, facial expression, voice expression, body posture, or eye tracking), measuring biosignals (e.g., blood pressure, heart rate, body temperature, sweat, or brain waves), identifying human (e.g., speech recognition, retinal recognition, facial recognition, fingerprint recognition, or EEG-based recognition), etc. Motion components 958 include acceleration sensor components (eg, accelerometers), gravity sensor components, rotation sensor components (eg, gyroscopes), and the like. Environmental components 960 include, for example, illuminance sensor components (e.g., a photometer), temperature sensor components (e.g., one or more thermometers to detect ambient temperature), humidity sensor components, pressure sensor components (e.g., barometers), acoustic sensor components (e.g., , one or more microphones to detect background noise), proximity sensor components (e.g., infrared sensors to detect nearby objects), gas sensors (e.g., machine smell detection sensors, to detect concentrations of harmful gases for safety, or to measure pollutants in the atmosphere gas detection sensors), or other components that can provide indications, measurements, or signals corresponding to the surrounding physical environment. Location component 962 includes a position sensor component (e.g., a global positioning system (GPS) receiver component), an altitude sensor component (e.g., an altimeter or a barometer that detects air pressure from which altitude can be derived), an orientation sensor component (e.g., a magnetic Count) etc.

Communications may be accomplished using a variety of techniques. I/O component 950 may include communication component 964 operable to couple machine 900 to network 980 or device 970 via coupling 982 and coupling 972, respectively. For example, communication component 964 includes a network interface component or another suitable device that interfaces with network 980 . In other examples, the communication component 964 includes a wired communication component, a wireless communication component, a cellular communication component, a near field communication (NFC) component, components (such as low energy), component, and other communicating components that provide communication via other modalities. Device 970 may be another machine or any of a variety of peripheral devices (eg, a peripheral device coupled via a Universal Serial Bus (USB)).

Additionally, in some implementations, the communicating component 964 detects an identifier or includes a component operable to detect an identifier. For example, the communication component 964 includes a radio frequency identification (RFID) tag reader component, an NFC smart label detection component, an optical reader component (e.g., for detecting one-dimensional barcodes such as Universal Product Code (UPC) barcodes), multi-dimensional barcodes (Optical sensors such as Quick Response (QR) Code, Aztec Code, Data Matrix, Data Word, MaxiCode, PDF417, Supercode, Uniform Commercial Code Reduced Space Symbol (UCC RSS)-2D barcode and other optical codes), acoustic detection Components (eg, a microphone that recognizes tagged audio signals) or any suitable combination thereof. Additionally, various information can be derived via the communication component 964, such as location via Internet Protocol (IP) geographic location, via The location of signal triangulation, the location of NFC beacon signals via detection may indicate a particular location, and the like.

Transmission medium

In various example embodiments, one or more portions of network 980 may be an ad hoc network, an intranet, an extranet, a virtual private network (VPN), a local area network (LAN), a wireless LAN (WLAN), a wide area network (WAN) , Wireless WAN (WWAN), Metropolitan Area Network (MAN), Internet, Part of the Internet, Part of the Public Switched Telephone Network (PSTN), Plain Old Telephone Service (POTS) Network, Cellular Telephone Network, Wireless Network, network, another type of network, or a combination of two or more such networks. For example, network 980, or a portion of network 980, may include a wireless or cellular network, and coupling 982 may be a Code Division Multiple Access (CDMA) connection, a Global System for Mobile Communications (GSM) connection, or other type of cellular or wireless coupling. In this example, coupling 982 may implement any of various types of data transmission technologies, such as single-carrier radio transmission technology (1xRTT), evolution data optimized (EVDO) technology, general packet radio service (GPRS) technology, Enhanced Data Rates for GSM Evolution (EDGE) technology, Third Generation Partnership Project (3GPP) including 3G, Fourth Generation Wireless (4G) networks, Universal Mobile Telecommunications System (UMTS), High Speed Packet Access (HSPA), Global Microwave Access Interoperability (WiMAX), Long Term Evolution (LTE) standards, other standards defined by various standards setting organizations, other long range protocols or other data transmission technologies.

In an example embodiment, a transport medium is used over network 980 via a network interface device (eg, a network interface component included in communication component 964) and utilizing one of a number of well-known transport protocols (eg, Hypertext Transfer Protocol (HTTP)). Either to send or receive instructions 916. Similarly, in other example embodiments, instructions 916 are sent to or received from device 970 via a coupling 972 (eg, a peer-to-peer coupling) using a transmission medium. The term "transmission medium" shall be taken to include any intangible medium capable of storing, encoding or carrying instructions 916 for execution by machine 900, and includes digital or analog communication signals or other intangible media used to facilitate communication of such software.

Additionally, a transmission medium or signal carrying machine-readable instructions comprises an embodiment of machine-readable medium 938 .

language

In this specification, plural instances may implement components, operations, or structures described as singular instances. Although various operations of one or more methods are illustrated and described as separate operations, one or more of the various operations may be performed concurrently, and the operations need not be performed in the order presented. Structure and functionality shown as separate components in example configurations may be implemented as a combined structure or component. Similarly, structures and functionality illustrated as a single component may be implemented as separate components. These and other variations, modifications, additions and improvements fall within the scope of the subject matter.

Although an overview of the inventive subject matter has been described with reference to specific example embodiments, various modifications and changes may be made to these embodiments without departing from the broader scope of the disclosed embodiments. These embodiments of the inventive subject matter may be referred to herein, individually or collectively, by the term "invention" for convenience only and are not intended to automatically limit the scope of the application to any single disclosure or inventive concept (if In fact more than one is disclosed).

The illustrated embodiments are described herein in sufficient detail to enable those skilled in the art to implement the disclosed teachings. Other embodiments may be utilized and derived from these embodiments, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the "details of embodiments" should not be taken in a limiting sense, and the scope of various embodiments is defined only by the appended claims, along with the full scope of equivalents to which the claims are entitled.

As used herein, the term "or" can be interpreted in an inclusive or exclusive sense. Furthermore, multiple instances may be provided for a resource, operation or structure described herein as a single instance. Additionally, the boundaries between the various resources, operations, modules, engines, and data stores are somewhat arbitrary, and certain operations are shown in the context of specific illustrative configurations. Other allocations of functionality are contemplated and may fall within the scope of the various embodiments of the present disclosure. In general, structures and functionality presented as separate resources in example configurations may be implemented as combined structures or resources. Similarly, structures and functionality presented as a single resource may be implemented as separate resources. These and other variations, modifications, additions and improvements fall within the scope of the embodiments of the present disclosure as represented by the appended claims. Accordingly, the specification and drawings are to be regarded as illustrative rather than restrictive.

The following enumerated examples define various example embodiments of the methods, machine-readable media, and systems (e.g., apparatuses) discussed herein:

Example 1 A system comprising:

The coarse category recognition module is configured as:

accessing a data set comprising categorical data having a plurality of fine-grained categories;

identifying a plurality of coarse categories, the number of coarse categories being less than the number of fine categories; and

For each fine category, an associated coarse category is determined;

The pre-training module is configured as:

Train an underlying convolutional neural network (CNN) to differentiate between coarse categories; and

training a fine CNN for each coarse category for discriminating between fine categories associated with the coarse category; and

The classification module, configured as:

receive requests to classify data;

Using the base CNN, determine a coarse category for the data;

determining a fine class of said data using the fine CNN of the determined coarse class; and

In response to the request, a fine category of the data is sent.

Example 2 The system according to Example 1, wherein the coarse category identification module is further configured to:

dividing the data set into a training set and a value set;

using the training set to train a first CNN model; and

generating a confusion matrix for the first CNN model using the set of values; wherein

Determining an associated coarse class for each fine class includes applying an affine propagation algorithm to the confusion matrix.

Example 3 The system according to example 1 or example 2, wherein the coarse category identification module is further configured to:

Obtain low-dimensional feature representations of fine-grained categories using Laplacian feature maps.

Example 4 The system of any suitable one of Examples 1 to 3, wherein the training module is configured to train a fine CNN for each coarse category using operations comprising:

using the training set to train a second CNN model;

generating a fine CNN for each coarse category based on the second CNN; and

A fine CNN for each coarse category is trained using a subset of the training set that does not include data with fine categories not associated with the coarse category.

Example 5. The apparatus of any suitable one of Examples 1 to 4, wherein:

The pre-training module is also configured to combine the CNN for distinguishing between coarse categories with each CNN for distinguishing between fine categories to form a Hierarchical Deep CNN (HD-CNN) ;as well as

The system also includes a fine-tuning module configured to fine-tune the HD-CNN.

Example 6 The system of example 5, wherein the fine-tuning module is configured to fine-tune the HD-CNN using operations comprising:

start said fine-tuning using a learning factor;

training the HD-CNN by iterating over a series of training batches using the learning factors;

After each iteration, comparing the training error of the training batch with a threshold;

determining, based on the comparison, that the training error for the training batch is below a threshold; and

The learning factor is modified in response to determining that the training error for the training batch is below a threshold.

Example 7 The system of example 5 or example 6, wherein the fine-tuning module is configured to fine-tune the HD-CNN using operations comprising:

A temporal sparsity element is applied in the evaluation of each of the described CNNs to differentiate between fine-grained categories.

Example 8 The system of any suitable one of Examples 1 to 7, wherein the data set comprising classified data comprises classified images.

Example 9 A method comprising:

accessing a data set comprising categorical data having a plurality of fine-grained categories;

identifying a plurality of coarse categories, the number of coarse categories being less than the number of fine categories;

For each fine category, an associated coarse category is determined;

training an underlying convolutional neural network (CNN) for distinguishing between coarse categories, the underlying CNN being implemented by the machine's processor; and

training a fine CNN for each coarse category for discriminating between fine categories associated with the coarse category;

receive requests to classify data;

Using the base CNN, determine a coarse category for the data;

determining a fine class of said data using the fine CNN of the determined coarse class; and

In response to the request, a fine category of the data is sent.

Example 10 The method according to Example 9, further comprising:

dividing the data set into a training set and a value set;

using the training set to train a first CNN model; and

generating a confusion matrix for the first CNN model using the set of values; wherein

Determining an associated coarse class for each fine class includes applying an affine propagation algorithm to the confusion matrix.

Example 11 The method according to Example 9 or Example 10, further comprising: using a Laplacian feature map to obtain a low-dimensional feature representation of a fine category.

Example 12 The method according to any one of Examples 9 to 11, wherein said training a fine CNN for each coarse category comprises:

using the training set to train a second CNN model;

generating a fine CNN for each coarse category based on the second CNN; and

A fine CNN for each coarse category is trained using a subset of the training set that does not include data with fine categories not associated with the coarse category.

Example 13. The method of any one of Examples 9-12, further comprising:

Combining the base CNN with each fine CNN to form a Hierarchical Deep CNN (HD-CNN); and

Fine-tune the HD-CNN.

Example 14 The method of any one of Examples 9 to 13, wherein fine-tuning the HD-CNN comprises:

start said fine-tuning using a learning factor;

training the HD-CNN by iterating over a series of training batches using the learning factors;

After each iteration, comparing the training error of the training batch with a threshold;

determining, based on the comparison, that the training error for the training batch is below a threshold; and

The learning factor is modified in response to determining that the training error for the training batch is below a threshold.

Example 15 The method according to any one of Examples 9 to 14, wherein the fine-tuning of the HD-CNN by the fine-tuning module includes:

A temporal sparsity element is applied in the evaluation of each of the described CNNs to differentiate between fine-grained categories.

Example 16. The method of claim 9, wherein the data set comprising classified data comprises classified images.

Example 17 A machine-readable medium carrying instructions executable by a processor of the machine to perform the method of any one of Examples 9-16.

Claims (18)

1. a kind of system, including：

Thick classification identification module, is configured as：

Access includes the data set of grouped data, and the grouped data has multiple fine classifications；

Multiple thick classifications are recognized, the quantity of the thick classification is less than the quantity of fine classification；And

For each fine classification, it is determined that associated thick classification；

Pre-training module, is configured as：

Train the basic convolutional neural networks (CNN) for being made a distinction between thick classification；And

The fine CNN of each thick classification of training, the fine CNN of the thick classification are used for associated with the thick classification fine Made a distinction between classification；And

Sort module, is configured as：

Receive the request classified to data；

Using the basic CNN, the thick classification of the data is determined；

Using the fine CNN of identified thick classification, the fine classification of the data is determined；And

In response to the request, the fine classification of the data is sent.

2. system according to claim 1, wherein, the thick classification identification module is additionally configured to：

The data set is divided into training set and value collects；

The first CNN models are trained using the training set；And

The confusion matrix of the first CNN models is generated using described value collection；Wherein

Determine that associated thick classification includes for each fine classification：Affine propagation algorithm is applied to the confusion matrix.

3. system according to claim 2, wherein, the thick classification identification module is additionally configured to：

The low-dimensional character representation of fine classification is obtained using laplacian eigenmaps.

4. system according to claim 2, wherein, the training module is configured with including the operation of the following To train the fine CNN of each thick classification：

The 2nd CNN models are trained using the training set；

The fine CNN of each thick classification is generated according to the 2nd CNN；And

Train the fine CNN of each thick classification using the subset of the training set, the subset do not include having with it is described thick The data of the unconnected fine classification of classification.

5. system according to claim 1, wherein,

The pre-training module is additionally configured to：By the CNN for being made a distinction between thick classification and in fine classification Between each CNN for making a distinction be combined, to form layering depth CNN (HD-CNN)；And

The system also includes fine setting module, and the fine setting module is configured as being finely adjusted the HD-CNN.

6. system according to claim 5, wherein, the fine setting module is configured with including the operation of the following To be finely adjusted to the HD-CNN：

Start the fine setting using Studying factors；

The HD-CNN is trained by using a series of training batches of the Studying factors iteration；

After each iteration, the training error of the training batch is compared with threshold value；

Based on it is described compare determine it is described training batch training error be less than threshold value；And

In response to determining that the training error of the training batch is less than threshold value, the Studying factors are changed.

7. system according to claim 5, wherein, the fine setting module is configured with including the operation of the following To be finely adjusted to the HD-CNN：

The sparse element of application time in the assessment to each CNN, to be made a distinction between fine classification.

8. system according to claim 1, wherein, including the data set of grouped data includes：Classification chart picture.

9. a kind of computer implemented method, including：

Access includes the data set of grouped data, and the grouped data has multiple fine classifications；

Multiple thick classifications are recognized, the quantity of the thick classification is less than the quantity of fine classification；

For each fine classification, it is determined that associated thick classification；

Train the basic convolutional neural networks (CNN) for being made a distinction between thick classification；

The fine CNN of each thick classification of training, the fine CNN of the thick classification are used for associated with the thick classification fine Made a distinction between classification；

Receive the request classified to data；

Using the basic CNN, the thick classification of the data is determined；

Using the fine CNN of identified thick classification, the fine classification of the data is determined；And

In response to the request, the fine classification of the data is sent.

10. method according to claim 9, in addition to：

The data set is divided into training set and value collects；

The first CNN models are trained using the training set；And

The confusion matrix of the first CNN models is generated using described value collection；Wherein

Determine that associated thick classification includes for each fine classification：Affine propagation algorithm is applied to the confusion matrix.

11. method according to claim 10, in addition to：The low-dimensional of fine classification is obtained using laplacian eigenmaps Character representation.

12. method according to claim 10, wherein, the fine CNN of each thick classification of training includes：

The 2nd CNN models are trained using the training set；

The fine CNN of each thick classification is generated according to the 2nd CNN；And

Train the fine CNN of each thick classification using the subset of the training set, the subset do not include having with it is described thick The data of the unconnected fine classification of classification.

13. method according to claim 9, in addition to：

The basic CNN is combined to form layering depth CNN (HD-CNN) with each fine CNN；And

The HD-CNN is finely adjusted.

14. method according to claim 13, wherein, the fine setting to the HD-CNN includes：

Start the fine setting using Studying factors；

The HD-CNN is trained by using a series of training batches of the Studying factors iteration；

After each iteration, the training error of the training batch is compared with threshold value；

Based on it is described compare determine it is described training batch training error be less than threshold value；And

In response to determining that the training error of the training batch is less than threshold value, the Studying factors are changed.

15. method according to claim 13, wherein, fine setting of the fine setting module to the HD-CNN includes：

The sparse element of application time in the assessment to each CNN, to be made a distinction between fine classification.

16. method according to claim 9, wherein, including the data set of grouped data includes：Classification chart picture.

17. a kind of non-transitory machine readable media, realize there is instruction thereon, the instruction can by machine computing device Include the operation of the following to perform：

Access includes the data set of grouped data, and the grouped data has multiple fine classifications；

Multiple thick classifications are recognized, the quantity of the thick classification is less than the quantity of fine classification；

For each fine classification, it is determined that associated thick classification；

Train the basic convolutional neural networks (CNN) for being made a distinction between thick classification, the CNN by machine processor Realize；

The fine CNN of each thick classification of training, the fine CNN of the thick classification are used for associated with the thick classification fine Made a distinction between classification；

Receive the request classified to data；

Using the basic CNN, the thick classification of the data is determined；

Using the fine CNN of identified thick classification, the fine classification of the data is determined；And

In response to the request, the fine classification of the data is sent.

18. a kind of machine readable media for carrying instruction, the instruction can be completed according to power by the computing device of machine Profit requires the method any one of 9 to 16.